LEARNING DISCOURSE STRUCTURE WITHOUT SEMANTICS: A STUDY ON TOPIC SEGMENTATION OBJECTIVES

Topic segmentation is commonly framed as a semantic similarity problem, yet it remains unclear whether optimizing for segmentation induces transferable sentence representations. In this work, we propose a differentiable, similarity-based framework that models topic segmentation as boundary detection over sentence-level similarity matrices. Our approach integrates a fully vectorized and differentiable segmentation head into a Transformer architecture, enabling end-to-end supervised training without assuming a predefined number of segments. Unlike classical matrix segmentation algorithms, the proposed head is designed for seamless integration with neural encoders and gradient-based optimization. We train the model from scratch on a newly curated Spanish Wikipedia dataset for topic segmentation, where sentence embeddings are transformed into cosine similarity matrices and processed to predict topic boundaries. The formulation explicitly decouples sentence encoding from segmentation while preserving full differentiability. Empirically, the model substantially outperforms zero-shot Transformer baselines and heuristic methods, achieving a WindowDiff score of 0.07 on the test set. Beyond performance, we conduct an in-depth analysis of training dynamics and representation transfer. Results indicate that, although the model effectively captures discourse-level structural shifts, the learned embeddings exhibit very limited transferability to semantic similarity tasks. These findings suggest that topic segmentation primarily exploits structural and discourse cues rather than inducing semantically aligned representation spaces.